Time, Time and Time Again: Disentangling Time Trends in Prescribing with Age Period Cohort Analysis, An Exemplar on the Annual Prevalence of Statin Prescribing for Males and Females in the United States (2000-2019)

Background: Cohort effects shape behavioral and physiological factors that may affect trends in prescribing and should be explored, yet are often ignored. Age period cohort (APC) analysis is one way to explore cohort effects, but has had little uptake in pharmacoepidemiology (PE).

Objectives: To demonstrate the utility of APC analysis in PE using an exemplar.

Methods: We conducted a repeated cross-sectional study on the annual prevalence of statin use using data on people aged 18-85 years in the United States who completed the Medical Expenditure Panel Survey Household Component (MEPS HC), 2000-2019. We defined statin use as having at least 1 statin prescription. We identified statin prescriptions linking MEPS HC Prescribed Medications data files to the Multum Lexicon database. We summarized observed annual prevalence of statin use by sex and period. We fitted a logistic hierarchical APC (HAPC) model: linear and quadratic age fixed effects, linear period fixed effect, and period and cohort cross-classified random effects. We chose the HAPC model because we were interested in describing nonlinearities in period and cohort effects that may be due to evolving clinical guidelines. We estimated predicted probabilities of age, period and cohort effects. We examined statistical significance of random effects using mixtures of Chi-squared distributions. We conducted analyses for each sex separately and did not account for the complex survey design.

Results: We included 129286 (47%) males and 147954 (53%) females aged 18-85 years. The prevalence of statin use was 9% in 2000 and 22% in 2019 for males; and 8% in 2000 and 18% in 2019 for females. We observed an increasing linear trend between 2000-2019 for both sexes. HAPC model results were similar by sex. The probability of statin use increased with age until plateauing for ages 80 years and over at around 0.40 for females and 0.45 for males, on average; as expected given the uncertain evidence on statin use in older people. The estimated period random effects suggested deviations from the linear trend between 2000-2003, with a lower probability of statin use relative to the linear increasing trajectory, on average. The estimated cohort random effects showed cohorts between 1916-1921 had a lower probability of statin use, beyond age and period effects, on average. This result may be in part due to differences in behavioral and physiological factors affecting statin use. There was strong evidence that the data were more compatible with the HAPC model relative to other models explored.

Conclusions: Accounting for cohort effects allow for a more accurate description of time trends. We demonstrated the use of one of many APC analytical approaches. The research question and substantive knowledge should inform the choice of the APC analytical approach.